AlGaN/GaN HEMT H₂ sensor with integrated Wheatstone bridge and on-chip microheater for 0.1-ppm detection Authors: Yu Yu, Yan Liu, Wenbin Ren, Yunpeng Liu, Yulong Zhang, Zihan Wang, Qiming Zhuo, Zewen Liu, Tao Wang, Jianwen Sun Published: 2026 ABSTRACT AbstractHydrogen (H₂) is a key clean-energy carrier in the transition to sustainable energy systems, but its extreme flammability and low ignition energy demand highly reliable leakage-monitoring technologies for safe deployment. This study presents a MEMS AlGaN/GaN high-electron-mobility transistor (HEMT) H₂ sensor that achieves high sensitivity and fast detection of trace hydrogen in air. The device implements a four-HEMT Wheatstone bridge—two Pd-gated sensing HEMTs and two passivated reference HEMTs—with an integrated microheater, all monolithically fabricated on a suspended GaN membrane. Pd-catalyzed H₂ dissociation and spillover lower the effective gate work function and modulate the Schottky barrier, and the differential readout suppresses common-mode drift. At 315 °C, the sensor achieves a limit of detection (LOD) of 0.1 ppm with a corresponding sensitivity of 1.48 mV/V/ppm. A combined device model and noise analysis further projects a detection limit approaching the ppb level. The baseline-corrected output is monotonic and well described by a four-parameter logistic model, thereby defining a practical, quantifiable working range of 0.1–1000 ppm (spanning four orders of magnitude). At 1000 ppm, the response and recovery times are 5.7 s and 26.2 s, respectively. The sensor exhibits high selectivity to H₂ with negligible cross-sensitivity to typical interferents (e.g., alcohols, SO₂, CH4, NH₃, H₂S, CO), along with outstanding repeatability over multiple cycles and long-term stability. This integrated, differential architecture is therefore suited for trace-level hydrogen-leak detection and early warning in applications such as semiconductor manufacturing and fuel-cell systems. Subject terms: Electrical and electronic engineering, Sensors Introduction Hydrogen (H₂) is widely recognized as a pivotal carbon-free energy carrier for sustainable energy systems; however, its widespread adoption faces critical safety challenges owing to its colorless and odorless nature, low minimum ignition energy, and wide flammability range. Consequently, reliable hydrogen leak-monitoring systems are essential. To ensure comprehensive safety—from early-stage leak detection to critical hazard prevention—gas sensors must meet stringent performance requirements. In practical deployments, an ultralow limit of detection (LOD) at the sub-ppm level is required because the natural atmospheric background concentration of hydrogen is ~0.6 ppm, enabling the identification of leaks at their very inception1. Beyond sensitivity, sensors must also provide a wide dynamic range that covers trace concentrations for early warning, as well as rapid response times—typically on the order of seconds—to trigger immediate safety protocols2. Furthermore, sensors must operate robustly in harsh environments, particularly at elevated temperatures exceeding 200 °C, where high-temperature hydrogen attack (HTHA) can induce material degradation such as blistering and cracking in conventional devices3. Over the past few decades, diverse sensing paradigms have been developed to address these safety imperatives. Non-electrical or RF-based detection methods have advanced significantly. Optical fiber sensors offer excellent immunity to electromagnetic interference and intrinsic safety, with recent studies demonstrating detection limits down to the sub-ppm level4,5. In parallel, surface acoustic wave (SAW) sensors—such as Pd/ZnO-functionalized MEMS devices—have demonstrated high sensitivity and fast response times suitable for wireless monitoring6. However, their widespread deployment is often constrained by bulky optical instrumentation (for fiber sensors) and susceptibility to environmental noise and vibration (for acoustic devices)4,6. In thermal sensing, conventional thermal conductivity detectors (TCDs) typically suffer from limited sensitivity (~2000 ppm)7,8, whereas recent advances in MEMS thermopiles have substantially lowered the detection limit. For example, Zhang et al. demonstrated a silicon-based differential thermopile sensor with a detection limit as low as ~1 ppm at 120 °C, representing state-of-the-art performance in calorimetric sensing9. Chemical sensors based on emerging nanomaterials have also attracted considerable attention, including Pd-decorated MXene (Ti₃C₂Tₓ) composites10 and precision-engineered Pd nanogap structures11, both enabling effective room-temperature detection. To further reduce power consumption, field-effect transistor (FET) platforms have been extensively explored, employing various channel materials including Si MOSFETs, organic FETs (OFETs), and low-dimensional nanostructures (e.g., carbon nanotubes, graphene)12–16. While these technologies have demonstrated specific merits such as ultralow power consumption (<50 µW)12 or sub-10 ppb detection13,14, they face distinct limitations: OFETs and nanostructured devices often struggle with environmental stability and fabrication reproducibility, respectively. More critically, a common thermal bottleneck persists: silicon-based devices (both thermopiles and FETs) face intrinsic material degradation above 150–200 °C9,12, while nanomaterial-based chemical sensors and SAW substrates are frequently plagued by oxidation or instability in harsh high-temperature environments. Consequently, a monolithic sensing platform that simultaneously achieves ultralow detection limits, rapid kinetics, and robust long-term reliability in high-temperature zones (>200 °C) remains elusive. From a deployment perspective, GaN HEMT sensors offer an all-electrical, low-footprint platform with inherent amplification and high-temperature robustness, thereby mitigating the intrinsically high power demand of calorimetric approaches and the high-temperature degradation commonly encountered in Si-based devices. Motivated by these practical requirements, gallium nitride (GaN) has emerged as an attractive material platform thanks to its exceptional chemical stability and high-temperature compatibility17. In particular, GaN-based high-electron-mobility transistors (HEMTs) enable a robust all-electrical transduction mechanism with intrinsic gain: surface-reaction-induced modulation of the gate potential is converted into pronounced drain-current changes through the device’s high transconductance18, yielding high signal-to-noise readout and straightforward circuit integration. Recent advances in gate functionalization on GaN HEMTs have unlocked remarkable performance, exemplified by Pt-decorated graphene-gated devices19 achieving a giant response of 1.6 × 107, WO₃-modified devices maintaining robust operation up to 250 °C20, and ZnO-nanoparticle/Pd dual-catalyst layers yielding enhanced sensitivity21,22. Representative H₂ sensors spanning these transduction routes, together with the device developed in this work for direct benchmarking, are compared in Table 1 to highlight the practical trade-offs among detection limit, response speed, operating temperature, and integration characteristics. Table 1Comparison of representative H₂ sensors based on different transduction mechanismsMechanismOperating temperatureDetection limitResponse timeRemarkRef.Optical fiberRT; practical use <60 °C0.3%90 sRequires optical interrogation; less amenable to monolithic on-chip integration5SAWRT/150 °C20,000 ppm—Limited trace sensitivity; often requires external RF interrogation6Thermal conductivity−15 to 84 °C2000 ppm—Wide range, but poor trace sensitivity7Thermopile MEMS120 °C~1 ppm1.9 s at 1000 ppmFast, but still thermal/catalytic-route dependent9CNT FETRT/~100 °C90 ppb (RT); 5 ppb ( ~ 100 °C)—Ultralow LOD, but manufacturability/stability challenges13CNT chemiresistorRT/100 °C890 ppb (RT); 89 ppb (100 °C)~7 s at ~311 ppmLow-LOD, but stability/robustness concerns14GaN HEMTRT0.5 ppm29 s at 10 ppmRobust platform, but drift-prone single-ended readout22GaN bridge HEMT315 °C0.1 ppm5.7 s at 1000 ppmDifferential readout with local microheating; monolithically integrated platformThis work Comparison of representative H₂ sensors based on different transduction mechanisms As summarized in Table 1, despite substantial progress across different H₂-sensing mechanisms, conventional GaN HEMT sensors still face critical architectural bottlenecks that hinder widespread industrial deployment. First, surface reaction kinetics on conventional planar devices are often diffusion-limited, resulting in sluggish response times typically ranging from 10 to 40 s23–26, falling short of rapid safety-shutdown requirements (<10 s). Second, achieving fast recovery and humidity tolerance typically requires elevated operating temperatures (200–300 °C). Most existing solutions rely on bulky external ceramic heaters27, which incur watt-level power consumption and hinder chip-scale integration. Although on-chip microheaters offer a pathway toward reduced system footprint, achieving thermally efficient platforms that sustain such temperatures without sacrificing thermal switching speed remains challenging. Furthermore, predominant single-ended metal–insulator–semiconductor high-electron-mobility transistor (MIS-HEMT) or Schottky configurations lack inherent common-mode rejection; consequently, threshold-voltage drifts arising from bias-temperature instability (BTI) and ambient thermal fluctuations can be indistinguishable from genuine gas signals17,18,28, leading to frequent false alarms. Therefore, unlocking the practical potential of GaN calls for a monolithic platform that synergizes differential drift cancellation with agile, thermally isolated local heating. In this study, we present a monolithically integrated GaN-on-Si sensing platform that addresses these critical trade-offs via a suspended differential-bridge architecture. To mitigate the baseline instability inherent to conventional single-ended devices, we implement a Wheatstone-bridge configuration that provides intrinsic common-mode rejection, thereby suppressing thermal drift and BTI-induced threshold-voltage shifts. In parallel, to reconcile elevated-temperature operation with system compactness, we design a suspended microheater with enhanced thermal isolation. This structural innovation confines heat generation to the active sensing region, eliminating the need for bulky external heating modules while enabling agile thermal modulation. Collectively, these architectural advances provide a practical pathway toward stable, fast, and highly sensitive hydrogen monitoring in harsh industrial environments. Results Sensor design and working principle As illustrated in Fig. 1a, the proposed device is a monolithically integrated hydrogen sensor whose core is a Wheatstone bridge composed of four AlGaN/GaN high-electron-mobility transistors (HEMTs) suspended on a circular GaN membrane. A serpentine platinum (Pt) microheater is integrated on the membrane to provide localized heating, which accelerates the adsorption–desorption kinetics of hydrogen and thereby enhances the sensing performance. The micromachined membrane offers excellent thermal isolation from the silicon substrate, effectively confining the generated heat within the active bridge area and minimizing lateral thermal losses. This design significantly reduces the operating power. Furthermore, the low thermal mass of the membrane enables rapid thermal response, short warm-up time, and precise temperature control. Architecture and working principle of the MEMS AlGaN/GaN HEMT H₂ sensor a Schematic of the suspended-membrane sensor incorporating two Pd-gated sensing HEMTs (left and right) and two passivated reference HEMTs (top and bottom) configured in a Wheatstone bridge, together with an integrated serpentine Pt microheater. b Simplified equivalent circuit model of the sensor. c Working-principle energy-band diagram of the sensor: Hydrogen atoms diffuse through the Pd gate and adsorb at the Pd/AlGaN interface, forming a dipole layer. This lowers the effective work function (from Φm₁ to Φm₂), acting as an equivalent positive gate bias that increases 2-DEG density and reduces channel resistance The left and right HEMTs, functionalized with 10-nm-thick palladium (Pd) catalytic gates, serve as the sensing elements. The Pd gate thickness is set to 10 nm as a practical engineering trade-off among response kinetics, film continuity, and structural stability, as motivated by prior studies. In principle, ultrathin Pd (≈2 nm) can shorten the diffusion path of dissociated hydrogen species and thus accelerate the response29. However, at the few-nanometer scale, Pd deposited on GaN may nucleate as islands or nanoparticles rather than fully coalescing into a continuous film, which can weaken gate continuity and reduce the stability of electrical modulation—an issue particularly critical for HEMT devices relying on a uniform gate30. Conversely, overly thick Pd layers (e.g., >20 nm) can undergo pronounced hydrogenation-induced lattice expansion during repeated absorption/desorption cycles, potentially introducing stress, deformation, and hysteresis, thereby compromising long-term stability and reversibility31. Considering these trade-offs, 10 nm avoids both extremes: it is thick enough to form a continuous conductive gate, while remaining thin enough to keep the diffusion barrier low and maintain rapid response kinetics. In contrast, the top and bottom HEMTs are covered with a thick metal/SiO₂ passivation stack, acting as on-chip references. These reference HEMTs are intentionally rendered hydrogen-insensitive but experience the same thermal and mechanical conditions as the sensing arms, allowing the differential Wheatstone configuration to effectively suppress common-mode drift and improve measurement stability. Pads 1–4 are designated for bridge excitation (Vin) and differential output (Vout) readout, while pads 8–9 are used to bias the microheater. Figure 1b presents the equivalent circuit model of the Wheatstone bridge. During practical sensing operations, the bridge is typically driven by a small excitation voltage (e.g., Vin ≤ 2 V) while the gate terminals remain floating. Under such small-bias conditions, the transistors operate deeply within the linear ohmic region, as experimentally verified by the bridge I–V characteristics (Supplementary Fig. S3). This physical behavior allows the four HEMTs to be accurately modeled as linear resistors. Based on the standard Wheatstone-bridge relation32, the differential output voltage is given by: Exposure to hydrogen reduces the resistances of the Pd-gated HEMTs (Rs1,Rs2), while the passivated reference HEMTs (Rr3,Rr4) remain unchanged, thereby unbalancing the bridge and generating a positive Vout signal. The superior sensitivity of the full-bridge design is evident when compared to a half-bridge configuration. Assuming an ideal initial balance where Rs1 = Rs2 = Rr3 = Rr4 = R0 in air, the resistance decrease for each sensing HEMT upon hydrogen exposure is defined as: so that the resulting small-signal outputs can be approximated as: This demonstrates that the full-bridge architecture provides approximately twice the small-signal response for an identical resistance change. In practice, fabrication asymmetries inevitably cause slight mismatches among the HEMTs, resulting in a static offset voltage, Voffset, measured as Vout in clean air at the operational temperature. To isolate the specific hydrogen response from this inherent imbalance and common-mode drift, we adopt a baseline-corrected metric: which is used as the sensor output in subsequent analysis. Figure 1c illustrates the operating mechanism of the Pd-gated sensing HEMT with and without hydrogen. In air, the Pd gate with a work function Φm1 forms a Schottky contact on AlGaN. The polarization-induced band bending at the AlGaN/GaN interface creates a 2-DEG. Meanwhile, the surface potential from the Pd contact further bends the AlGaN bands upward, partially depleting the channel and establishing a baseline 2-DEG density, which results in the baseline current. In H₂, hydrogen molecules dissociate on the Pd surface, and the generated hydrogen atoms (H) diffuse through the Pd layer and adsorb at the Pd/AlGaN interface33,34. These adsorbed hydrogen atoms form a dipole layer at the interface, which effectively lowers the work function of the Pd gate from Φm1 to Φm2 and is electrostatically equivalent to applying a positive gate bias35,36. This causes the conduction band edge to move downward, increasing the 2-DEG density. As a result, the channel resistance of the sensing HEMT decreases, leading to an increase in the differential bridge output. In contrast, the reference HEMTs are passivated with a metal/SiO₂ stack, which not only physically blocks hydrogen from reaching the gate but also remains chemically inert, thereby preserving their original band structure and electrical characteristics. Fabrication and structural characterization Figure 2 summarizes the complete micro/nanofabrication process. The sensors are fabricated at the wafer level on a 6-inch AlGaN/GaN-on-Si epitaxial wafer with the following layer stack from bottom to top: high-resistivity 〈111〉 Si (675 μm), a 1600 nm GaN buffer layer whose bottom includes a 100 nm AlN nucleation layer, an unintentionally doped GaN channel (300 nm), an AlN interlayer (0.8 nm), and an Al₀.₂₆Ga₀.₇₄N barrier (27 nm) (Fig. 2a). Mesa isolation is first defined by etching 100 nm into the heterostructure using a BCl₃/Cl₂/Ar plasma (Fig. 2b). Ohmic contacts are then formed by electron-beam evaporation of a Ti/Al/Ti/Au (20/110/40/50 nm) metal stack, followed by lift-off patterning and rapid thermal annealing (RTA) at 840 °C for 45 s in N₂ ambient. Next, a 300 nm SiO₂ passivation layer is deposited by plasma-enhanced chemical vapor deposition (PECVD) (Fig. 2c). Micro/nanofabrication process flow for the MEMS AlGaN/GaN HEMT H₂ sensor. a Layer structure of the AlGaN/GaN-on-Si epitaxial wafer. b Mesa isolation defined by dry etching. c Ohmic contact formation and passivation. d Etching of Ohmic contact and gate windows. e Formation of Wheatstone-bridge interconnects, serpentine Pt microheater, and metal shields for the reference HEMT gates. f Second passivation layer deposition. g Opening of sensing-gate areas and bond pads. h Deposition of Cr/Pd catalytic gates for the sensing HEMTs. i Backside wafer thinning and DRIE of the Si substrate to form the circular suspended membrane To open the source/drain and gate windows, 250 nm of SiO₂ is first removed by CF₄ plasma etching, and the remaining 50 nm is gently removed using buffered oxide etch (BOE) to mitigate plasma-induced damage to the underlying AlGaN surface (Fig. 2d). Subsequently, the Wheatstone-bridge interconnects, serpentine microheater, and metal shields for the reference HEMT gates are formed by depositing and patterning a Ti/Pt/Ti (10/300/10 nm) stack (Fig. 2e). Another 300 nm SiO₂ passivation layer is then deposited by PECVD (Fig. 2f). The sensing gates and bond pads are exposed by partially etching the SiO₂ with CF₄ plasma, followed by BOE removal of the final 50 nm (Fig. 2g). A Cr/Pd (2/10 nm) bilayer is sputtered and patterned via lift-off to form the Pd catalytic gates of the sensing HEMTs (Fig. 2h). Here, the ultrathin Cr layer serves as an adhesion layer to improve Pd film continuity and wafer-level process yield on the AlGaN surface, while its 2 nm thickness minimizes any impact on hydrogen access to the Pd catalyst. The Pd thickness (10 nm) is a nominal value calibrated from the sputtering deposition rate and excludes the 2 nm Cr adhesion layer. Finally, backside mechanical grinding thins the wafer to 400 μm. The silicon substrate beneath the active area is then selectively removed by deep reactive ion etching (DRIE) from the backside to create a 600 μm-diameter circular suspended membrane, which provides excellent thermal isolation (Fig. 2i). The MEMS AlGaN/GaN HEMT H₂ sensor is systematically characterized using optical microscopy, focused-ion-beam (FIB) cross-sectioning, and transmission electron microscopy (TEM) combined with energy-dispersive X-ray spectroscopy (EDS), as summarized in Fig. 3. Optical micrographs show the overall die size of 1 mm × 1 mm, a magnified view of the Wheatstone-bridge region, and the wire-bonded connections to a ceramic carrier (Fig. 3a–c). In Fig. 3b, the Pd-coated sensing HEMT gates and the metal/SiO₂-shielded reference HEMT gates exhibit clearly distinct optical contrast, reflecting their different gate stacks and being consistent with the FIB/TEM/EDS analyses discussed below. Cross-sectional FIB imaging provides a clear view of a sensing HEMT on the suspended membrane (Fig. 3d). Most importantly, high-resolution TEM imaging unambiguously resolves the detailed heterostructure of the active region, including the Pd catalytic gate, the AlGaN barrier layer, and, critically, the 2-DEG channel confined at the AlGaN/GaN heterointerface (Fig. 3e, f). Cross-sectional TEM also confirms a continuous Pd film, while slight local thickness variations are observed, which can be attributed to surface roughness and the specific TEM cross-section location. Plan-view EDS mapping around the sensing gate and cross-sectional elemental analysis across the GaN epitaxial stack further confirm uniform deposition of the Pd thin film on the AlGaN/GaN surface (Fig. 3g–j). Structural characterization of the MEMS AlGaN/GaN HEMT H₂ sensor. a Optical micrograph of the 1 mm × 1 mm AlGaN/GaN HEMT H₂ sensor die. b Magnified view of the Wheatstone-bridge region, clearly showing the Pd-gated sensing HEMTs and the metal/SiO₂-shielded reference HEMTs. c Photograph of the sensor wire-bonded to a ceramic package. d Cross-sectional FIB image of a sensing HEMT located on the suspended GaN membrane. e Cross-sectional view of the gate region of the sensing HEMT. f Enlarged TEM view of the Pd gate/AlGaN barrier/GaN channel stack, highlighting the Pd/AlGaN contact and the AlGaN/GaN heterointerface that hosts the 2-DEG. g–j Plan-view and cross-sectional EDS elemental maps confirming uniform deposition of the Pd thin film on the AlGaN/GaN surface Microheater characterization and temperature calibration Figure 4 characterizes the on-chip serpentine Pt microheater and establishes the electrical-to-temperature calibration used throughout this work. On a probe station, the heater is biased from 1 to 7 V. For each bias, we record the membrane temperature by infrared thermography and the steady-state heater resistance Rh after thermal stabilization. The chip temperature T increases monotonically with heater voltage Vh, consistent with Joule heating (Fig. 4a). Figure 4b shows representative infrared thermographs at 5 V and 7 V, where the hot region is confined to the suspended GaN membrane while the surrounding frame remains comparatively cool; the enlarged insets (Fig. 4e, f) provide zoomed-in views of the membrane region to more clearly visualize this localized heating behavior. Thermal characterization and electrical-to-temperature calibration of the suspended-membrane platform. a Membrane temperature T as a function of heater voltage Vh extracted from infrared thermography. b Representative infrared thermographs at 5 V and 7 V bias, demonstrating successful heat localization within the suspended membrane; e, f show enlarged views of the membrane region for the corresponding 5 V and 7 V cases. c Self-sensing calibration curve linking heater resistance Rh to temperature T, obtained by fitting the exponential model in Eq. (6). d Operating-temperature map for different combinations of Vh and bridge excitation Vin, used as the reference for all subsequent measurements As shown in Fig. 4c, using the seven (Rh, T) pairs, we construct a self-sensing resistance–temperature calibration curve. The relationship between heater resistance and chip temperature is empirically fitted from our calibration data as: which yields fit parameters a ≈ 66.18 Ω and b ≈ 1.86 × 10−3 °C−1 with a coefficient of determination R2 = 0.99. This empirical exponential dependence, which deviates from the near-linear or second-order behavior of bulk platinum RTDs described by the Callendar–Van Dusen equation37–39, is commonly observed in Pt thin films and is attributed to microstructure-dependent electron scattering at grain boundaries and surfaces20,40, as well as film–substrate thermal strain and residual stress introduced during microfabrication33,37,41. Nevertheless, the high coefficient of determination confirms its exceptional reliability for in situ thermal monitoring in this study. This one-to-one mapping enables in situ temperature readout via: This relationship is applied strictly within the calibrated range. During sensing experiments, we apply different combinations of heater voltage (Vh) and bridge excitation (Vin), as summarized on the x-axis labels of Fig. 4d. To account for the bridge-induced alteration of the local thermal load, the working temperature for each operating point is consistently determined by in situ reading of Rh and conversion through the calibration of Fig. 4c. This procedure eliminates emissivity-related uncertainties and provides a consistent chip temperature for every (Vh, Vin) condition, as visualized by the derived temperature scale in Fig. 4d. This consistent thermal basis underpins all subsequent comparisons of sensitivity, response/recovery time, and power consumption. Hydrogen sensing performance of the MEMS AlGaN/GaN HEMT H₂ sensor Figure 5 surveys the operating space at 1000 ppm H₂ (test setup in Supplementary Fig. S1a) and guides the choice of bias conditions based on four key performance metrics. Six sensors are tested in total. Because the gas concentration in the test system does not reach the target value instantaneously, the extracted power consumption and baseline-corrected output ΔV are reliable, whereas the response time (τr, 0–90%) and recovery time (τd, 100–10%) are approximate, reflecting the gradual change in gas concentration. Nonetheless, their trends are robust and can be used to guide bias selection. The corrected output ΔV, power, response time, and recovery time are mapped as functions of the heater voltage Vh and bridge excitation Vin; error bars denote die-to-die variation (Fig. 5a–d). The red crosses (×) in the response time and recovery time maps (Fig. 5c, d) indicate operating points where no valid sensor reading is obtained. General performance and bias selection of the MEMS AlGaN/GaN HEMT H₂ sensor at 1000 ppm H₂. a Baseline-corrected output ΔV as a function of heater voltage and bridge excitation Vin. b Total power consumption corresponding to each (Vh, Vin) operating point. c Response time (τr, 0–90%) versus Vh and Vin. d Recovery time (τd, 100–10%) versus Vh and Vin As Vh increases, the sensor transitions through three distinct performance regimes governed by the thermal energy supplied to the sensing membrane. At low Vh (≤3 V), the device operates in an under-heated regime. In this regime, the membrane temperature is likely inadequate to sufficiently activate the surface processes required for efficient hydrogen dissociation and subsequent desorption, resulting in suppressed apparent kinetics. This is consistent with the observed small signal (ΔV) and prolonged response and recovery times. When Vh is raised to the optimal range of 4–5 V, the device enters a high-performance sensing regime: the increased thermal budget more effectively activates the catalytic surface processes, while the total power increases predominantly with Vh (Fig. 5b), confirming that the microheater dominates the power budget. As a result, both the response time τr and the recovery time τd are generally shortened. Further pushing the heater to Vh = 6 V enlarges ΔV but incurs high total power and drives the membrane temperature beyond 350 °C, a range where conventional Ti/Al-based ohmic contacts are known to exhibit long-term drift and degradation under continuous operation42,43. These conditions can compromise the long-term reliability of the device. The dependence of the sensor performance on Vin further explains the observed error bars. When Vin≥ 2.5 V, small branch mismatches and bias-dependent resistances are amplified by the nonlinearity of an imperfectly balanced Wheatstone bridge (as corroborated by the I–V curve deviation in Supplementary Fig. S3), so the spread grows markedly; by contrast, for Vin ≤ 2 V the bridge operates closer to the linear small-signal regime and the spread is much smaller, albeit at the cost of reduced ΔV if Vin < 2 V. The trends revealed in Fig. 5 are critical for defining the sensor’s practical operating window. Balancing ΔV, power, response time, recovery time, and uniformity, we use Vh = 5 V and Vin = 2 V in subsequent experiments, corresponding to a target operating temperature of 315 °C and a total power consumption of 219 mW. This on-chip heating capability enables a shift away from prior high-performance GaN sensors that rely on bulky external laboratory heating systems (e.g., hot chucks) to reach their operating temperatures20,36, thereby supporting portable, low-power deployment. All subsequent measurements are performed on a representative sensor with median performance. Figure 6 comprehensively evaluates the hydrogen sensing performance across a wide concentration range from 0.1 to 6000 ppm. As shown in Fig. 6a–e, the corrected output ΔV increases monotonically with concentration and begins to approach saturation above ~1000 ppm and reaches a near-saturation plateau beyond ~3000 ppm. Here, the apparent saturation refers to the electrical transduction (ΔV) approaching a plateau, rather than Pd ceasing hydrogen uptake. In Pd-gated AlGaN/GaN HEMTs, the output is governed by the hydrogen-induced Pd/AlGaN interfacial dipole (effective work-function/Schottky-barrier shift) that modulates the 2-DEG; as the interfacial coverage approaches saturation at high H₂ concentration, the incremental barrier/2-DEG modulation diminishes, yielding a sub-linear response and a near-plateau44. Although the sensor remains responsive up to 6000 ppm, the incremental change in ΔV above ~ 3000 ppm is <0.2 mV, leading to markedly degraded concentration resolution near saturation; therefore, in the following analysis, data above 3000 ppm are not used for quantitative calibration fitting and are discussed only as high-concentration saturation behavior. Notably, at 0.1 ppm, the output is ~0.35 mV above the baseline, indicating clear potential for the detection of even lower concentrations. H2 sensing performance of the MEMS AlGaN/GaN HEMT H2 sensor. a–e The corrected output ΔV at various H₂ concentrations from 0.1 to 6000 ppm, showing a monotonic increase and saturation at high concentrations. f Steady-state ΔV versus concentration and corresponding four-parameter logistic fit, demonstrating excellent agreement between experiment and model. g Variation of relative response ΔVrel with concentration in the 0–0.1 ppm range, with 3σ and 10σ threshold lines marking the limit of detection (LOD) and limit of quantitation (LOQ). h Sensitivity S as a function of H₂ concentration, highlighting the higher sensitivity achieved at lower concentrations To enable quantitative readout, the steady-state ΔV values at each concentration are compiled into a calibration curve (Fig. 6f). Motivated by the saturating concentration dependence expected from Langmuir-type surface adsorption kinetics44,45, the full dataset over 0.1–3000 ppm is fitted using the following empirical four-parameter logistic model: with an excellent coefficient of determination (R2 = 0.99) and representative parameters A1 ≈ −0.42 mV, A2 ≈ 69.14 mV, x0 ≈ 75.32 ppm, and p ≈ 0.79. This empirical functional form effectively captures the observed saturation behavior of the Pd-gated response: the response rises rapidly at low concentrations due to abundant vacant sites and saturates at high concentrations as sites become occupied, mirroring the physical process of surface adsorption and 2-DEG modulation in the AlGaN/GaN heterostructure. In this framework, C denotes the H2 concentration in ppm; the parameter A1 represents the baseline output, corresponding to the sensor’s response in pure air, influenced by electronic noise and offset. A2 denotes the saturation output, indicating the maximum response achievable at high hydrogen concentrations, limited by the Pd catalytic capacity and 2-DEG density saturation. The parameter x0 is the inflection point concentration, where the response reaches half-saturation, reflecting the adsorption affinity on Pd surfaces. The parameter p controls the curve steepness around x0, with p < 1 indicating a gradual transition, possibly due to surface heterogeneity or cooperative adsorption effects. Although this global fit confirms a monotonic and one-to-one mapping across the full measured range, it does not imply uniform concentration resolution over the entire 0.1–3000 ppm window. To further evaluate the noise-limited quantification capability, we perform a detailed analysis of the sensor’s baseline stability. As illustrated in Supplementary Fig. S2, a 60-second steady-state segment is extracted from the recovery phase of the transient response curve (originally shown in Fig. 6e, time interval: 4123–4183 s). The signal is recorded at a sampling frequency of 6.7 Hz. The root-mean-square (RMS) noise deviation for this interval is calculated to be σ = 0.0103 mV. Although the sensor still maintains an experimentally detectable signal near the upper end of the test window (for example, from 2500 to 3000 ppm, ΔV increases by ~0.26 mV, corresponding to a signal-to-noise ratio (SNR) of >25), its concentration resolution is inevitably affected as the response gradually approaches saturation. Because the response slope becomes flatter at high concentrations, distinguishing small concentration increments becomes increasingly difficult in practical operation. Therefore, in order to fully leverage the sensor’s advantages in ultra-low limit of detection (LOD) and early leak warning, we define the 0.1–1000 ppm range as its primary high-resolution quantitative working window, whereas the response above 1000 ppm is interpreted mainly as a threshold warning signal for high-concentration hydrogen environments. On the basis of the above noise analysis, we further estimate the sensor’s theoretical detection capability by combining the fitted calibration model with the measured baseline fluctuation. To reference the response to the blank and remove the extrapolated baseline offset A1, we define the blank-normalized response: The inset of Fig. 6f (also shown separately as Fig. 6g) plots ΔVrel versus concentration in the 0–0.1 ppm range. Substituting y=A1+kσ into Eq. (8) and inverting directly yields the corresponding concentration limits for k = 3 and k = 10: we obtain a theoretical limit of detection CLOD = 4.60 ppb (95% confidence interval: 4.17–5.11 ppb) and a limit of quantitation (LOQ) CLOQ = 20.93 ppb (95% confidence interval: 18.98–23.26 ppb). Although our experimental lower bound is 0.1 ppm, this combined model-and-noise analysis indicates ppb-level capability, with SNR = ΔVrel/3σ ≈ 11.5 at 0.1 ppm, leaving substantial headroom for further down-range measurements. Furthermore, the sensor’s sensitivity S is analyzed as a function of concentration (Fig. 6h). The sensitivity is defined as: At 0.1 ppm, the sensitivity reaches an exceptionally high value of 1.48 mV/V/ppm, underscoring the sensor’s superior response at trace levels. The sensitivity is highest at low hydrogen concentrations and gradually declines as the concentration increases. Physically, when H₂ is scarce, abundant vacant Pd sites are available, so each additional ppm produces a relatively large increase in surface coverage, which strongly lowers the effective gate work function, reduces the Schottky barrier, and boosts the 2-DEG density in the AlGaN/GaN channel. At higher concentrations, those surface sites approach saturation, so the incremental electrical change per ppm becomes smaller and the sensitivity drops. This greater sensitivity at lower concentrations highlights the sensor’s particular suitability for trace-level detection applications. Additionally, we evaluate the repeatability, long-term reliability, intrinsic response/recovery times, and selectivity of the MEMS AlGaN/GaN HEMT H2 sensor. As shown in Fig. 7a, the device undergoes 30 cycles at 1000 ppm H₂; the corrected output is normalized to its 30-cycle mean of 59.95 mV, yielding a cycle-to-cycle variation from −0.8% to +0.6% and showing negligible waveform distortion, thereby demonstrating excellent repeatability. Repeatability, stability, intrinsic dynamics, and selectivity of the MEMS AlGaN/GaN HEMT H2 sensor. a Cyclic response of the sensor over 30 exposure–recovery cycles at 1000 ppm H₂ (ΔV shown on the left axis; the right axis shows the normalized cycle-to-cycle variation), showing minimal fluctuation and excellent repeatability. b Long-term stability test over 60 days, with measurements taken every 6 days and normalized to the initial response. c Intrinsic response and recovery transients at 1000 ppm H₂ obtained using a step-change (manual-switching) setup. d Response time at representative H₂ concentrations over the 0.1–800 ppm range. e Sensor responses to various combustible gases, volatile organic compounds (VOCs), and other harmful gases (each at 1000 ppm), compared with the response to H₂, demonstrating high chemical selectivity To assess its operational durability, a 60-day stability study is then conducted. Measurements are taken every 6 days and normalized to the day-0 value. The normalized output fluctuates within a narrow range of −3.0% to +3.2%, demonstrating robust long-term stability of the Pd-gated HEMT bridge at the operating temperature (Fig. 7b). Furthermore, because the previous test setup (Supplementary Fig. S1a) ramps the chamber concentration slowly and cannot reveal the intrinsic kinetics, we employ a manual switching setup to generate a sharp step change from air to 1000 ppm H₂ (Supplementary Fig. S1b). This method reveals a response time τr of 5.7 s and a recovery time τd of 26.2 s, confirming second-scale dynamics (Fig. 7c). In addition, we compile the response time at several representative concentrations over the 0.1–800 ppm range (Fig. 7d). Overall, the response time decreases with increasing H2 concentration; this survey provides an intuitive overview of the response-speed evolution from trace to high concentrations. Mechanistically, higher H₂ concentrations lead to a higher molecular flux on the catalytic Pd surface. The increased H₂ partial pressure provides a stronger kinetic driving force for hydrogen adsorption/dissociation and the subsequent interfacial potential modulation at the Pd/AlGaN gate, thereby accelerating the electrical response and shortening the response time. Finally, selectivity is examined by exposing the sensor to 1000 ppm of various interfering gases, including methanol, ethanol, ethylene glycol, n-butanol, SO₂, CH4, NH₃, H₂S, and CO. The response to H₂ greatly exceeds those to the other gases under the same conditions, confirming excellent chemical selectivity (Fig. 7e). This exceptional specificity toward H₂ against the aforementioned VOCs and inorganic gases is governed by a three-level catalytic–transport–transduction selectivity mechanism. First, Pd is a highly selective catalyst for dissociative adsorption of H₂, readily generating atomic hydrogen. In contrast, the interfering molecules (e.g., CH₄ and alcohols) are chemically more stable and their dissociation on Pd is kinetically unfavorable under the same operating conditions, so they do not produce an equivalent population of atomic hydrogen. Second, the dense Pd gate effectively acts as a hydrogen-permeable layer: dissociated H atoms can diffuse through Pd and reach the buried Pd/AlGaN interface, whereas the transport of larger molecular species (and their fragments) through the Pd film to this buried interface is negligible on the experimental timescale. Finally, because the HEMT response is governed by the dipole layer formed at the Pd/AlGaN interface, species that cannot generate and deliver atomic hydrogen to this interface induce only weak interfacial potential modulation and thus a negligible 2-DEG change. Overall, the proposed MEMS AlGaN/GaN HEMT H₂ sensor achieves a 0.1-ppm detection limit and retains second-scale kinetics at 1000 ppm. Across 0.1–3000 ppm, the corrected output is monotonic and well described by a four-parameter logistic model. Compared with reported AlGaN/GaN HEMT sensors23,24,26, MOS-structure sensors25,27,46, and Schottky-diode sensors35,47,48, as shown in Table 2, the proposed sensor achieves an LOD that is typically 1–2 orders of magnitude lower. While direct kinetic comparisons across the literature are inherently complex due to significant variations in operating temperatures, heating configurations, and target gas concentrations, our sensor achieves highly competitive, seconds-level response and recovery times. Notably, unlike many literature counterparts that report rapid kinetics under extreme hydrogen concentrations (e.g., 10,000 ppm) or rely on externally heated chambers or hot chucks (which artificially accelerate reactions by pre-heating the ambient gas), our device achieves these rapid kinetics at a moderate 1000 ppm through an ultra-compact, localized microheater. Collectively, these results demonstrate a superior overall balance among detection limit, thermal management efficiency, and response kinetics relative to previously reported counterparts. Table 2Comparison of key performance of three types of GaN-based H2 sensorsGaN-based sensor typeTemp. & heatingLOD (ppm)τr @ C (ppm)τd @ C (ppm)Refs.HEMT structure315 °C (MH)0.15.7 s @ 1000 ppm26.2 s @ 1000 ppmThis workHEMT structureRT (−)19 s @ 2000 ppm31 s @ 2000 ppm23HEMT structure127 °C (HC)5~ 12 s @ 1000 ppm~22 s @ 1000 ppm24HEMT structureRT (−)19 s @ 10000 ppm15 s @ 10000 ppm26Schottky diode177 °C (HC)1414 s @ 9970 ppm9 s @ 9970 ppm48Schottky diode127 °C (HC)1~12 s @ 1000 ppm~16 s @ 1000 ppm35Schottky diode160 °C (HC)0.821 s @ 10000 ppm8 s @ 10000 ppm47MOS structure150 °C (HC)1~7 s @ 1000 ppm~8 s @ 1000 ppm25MOS structure67 °C (HC)5~30 s @ 10000 ppm~7 s @ 10000 ppm27MOS structureRT (−)0.1120.4 s @ 1000 ppm25.4 s @ 1000 ppm46LOD limit of detection, τr response time, τd recovery time, MH Monolithically integrated microheater, HC External hot-chuck or globally heated chamber, RT Room temperature, – Room-temperature testing without active heating Comparison of key performance of three types of GaN-based H2 sensors LOD limit of detection, τr response time, τd recovery time, MH Monolithically integrated microheater, HC External hot-chuck or globally heated chamber, RT Room temperature, – Room-temperature testing without active heating Conclusion In summary, this work presents a MEMS AlGaN/GaN HEMT H2 sensor featuring a monolithically integrated Wheatstone bridge structure with an embedded microheater. The bridge combines Pd-gated sensing HEMTs with passivated reference HEMTs, enabling accurate differential readout of hydrogen-induced changes in channel resistance. Leveraging a suspended membrane with on-chip heating, the device provides excellent thermal isolation and precise temperature control, which markedly improve reaction kinetics and measurement stability. The sensor delivers a well-balanced overall performance: a 0.1-ppm LOD with a corresponding sensitivity of 1.48 mV/V/ppm at this concentration; a primary high-resolution quantitative working range of 0.1–1000 ppm, spanning four orders of magnitude, together with rapid response/recovery times of 5.7 s and 26.2 s, respectively. Furthermore, the sensor exhibits excellent selectivity to H₂, remarkable repeatability, and long-term stability. These attributes highlight the MEMS AlGaN/GaN HEMT H₂ sensor as a strong candidate for a wide range of applications, particularly trace-level detection and hydrogen-leak early-warning systems. Building on this system-ready monolithic platform, future work will focus on integrating the proposed sensor with customized CMOS readout circuits to construct a complete, fully packaged portable microsystem, thereby further validating its effectiveness in practical real-world deployments. Supplementary information Supplemental Material Acknowledgements This work was supported by Department of Science and Technology of Zhejiang Province (2023C01182). Author contributions Y.Y. and Y.L. contributed equally to this work as co-first authors. Y.L., T.W., and J.S. conceived the experiments, provided support, and handled the correspondence for this article; Y.Y. carried out the process flow and wafer fabrication, and performed the majority of the data acquisition, analysis, processing, and manuscript preparation; W.R., Yunpeng Liu, Z.W., Q.Z. contributed to parts of the analysis and discussion; Z.L., Y.Z. contributed to parts of the analysis and polishing of the article. Conflict of interest The authors declare no competing interests. Footnotes These authors contributed equally: Yu Yu, Yan Liu. Contributor Information Yan Liu, Email: liuyan2010@zju.edu.cn. Tao Wang, Email: twang@zju.edu.cn. Jianwen Sun, Email: sunjianwen@megasensing.cn. Supplementary information The online version contains supplementary material available at 10.1038/s41378-026-01366-2. References Associated Data Supplementary Materials Supplemental Material